Data Structures for Irregular Applications

Parallelization of any large application can be a diicult task, but when the application contains irregular patterns of communication and control, the parallelization eeort is higher and the likelihood of producing an eecient implementation is lower. The kinds of data structures that appear in irregular applications, for example, trees, graphs, and sets, do not have simple mappings onto distributed memory machines. We are building a library of such distributed data structures that use a combination of replication and partitioning to achieve high performance. Operations on these structures cannot be eeciently implemented as atomic operations, because of the latency of inter-processor communication. We propose a relaxed consistency model for these data structures, which is analogous to a weak consistency model on shared memory. This allows for clean, simple interfaces on the objects, but admits low latency, high throughput implementations. We demonstrate these ideas using a few data structure examples, each of which is being used in at least one application.